wave13: 514/239 — flink/nifi/artemis + K8s/cilium/linkerd2 + ES/OpenSearch/Solr + hadoop/hbase/spark

defects/artemis/patch/artemis-0001-bindingsimpl-routefromcluster-hashset.md

@@ -0,0 +1,91 @@
+# artemis-0001: BindingsImpl routeFromCluster O(R×A) → O(R+A)
+
+## Location
+`artemis-server/src/main/java/org/apache/activemq/artemis/core/postoffice/impl/BindingsImpl.java`
+Lines 630–664 (`routeFromCluster`)
+
+## Severity
+HIGH — called on every clustered message that requires selective ACK routing
+
+## Description
+`routeFromCluster` decodes two byte arrays: `ids` (all binding IDs to route to) and
+`idsToAck` (subset that require ACK). It builds `idsToAckList` as an `ArrayList<Long>`,
+then iterates over `ids` calling `idsToAckList.contains(bindingID)` for each entry.
+
+For R route targets and A ACK targets, this is O(R×A). Since this method is called
+on the hot message routing path in clustered deployments, it compounds per message.
+
+## Root Cause
+```java
+private void routeFromCluster(final Message message, final RoutingContext context,
+                               final byte[] ids) throws Exception {
+    byte[] idsToAck = (byte[]) message.removeProperty(Message.HDR_ROUTE_TO_ACK_IDS);
+    List<Long> idsToAckList = new ArrayList<>();          // ← ArrayList
+
+    if (idsToAck != null) {
+        ByteBuffer buff = ByteBuffer.wrap(idsToAck);
+        while (buff.hasRemaining()) {
+            idsToAckList.add(buff.getLong());             // populate O(A)
+        }
+    }
+
+    ByteBuffer buff = ByteBuffer.wrap(ids);
+    while (buff.hasRemaining()) {                         // O(R) loop
+        long bindingID = buff.getLong();
+        Binding binding = bindingsIdMap.get(bindingID);
+        if (binding != null) {
+            if (idsToAckList.contains(bindingID)) {      // O(A) scan → O(R×A) total
+                binding.routeWithAck(message, context);
+            } else {
+                binding.route(message, context);
+            }
+        }
+    }
+}
+```
+
+## Fix
+Use a `HashSet<Long>` instead of `ArrayList<Long>` for `idsToAckSet`:
+
+```java
+private void routeFromCluster(final Message message, final RoutingContext context,
+                               final byte[] ids) throws Exception {
+    byte[] idsToAck = (byte[]) message.removeProperty(Message.HDR_ROUTE_TO_ACK_IDS);
+    Set<Long> idsToAckSet = new HashSet<>();              // ← HashSet
+
+    if (idsToAck != null) {
+        ByteBuffer buff = ByteBuffer.wrap(idsToAck);
+        while (buff.hasRemaining()) {
+            idsToAckSet.add(buff.getLong());              // populate O(A)
+        }
+    }
+
+    ByteBuffer buff = ByteBuffer.wrap(ids);
+    while (buff.hasRemaining()) {                         // O(R) loop
+        long bindingID = buff.getLong();
+        Binding binding = bindingsIdMap.get(bindingID);
+        if (binding != null) {
+            if (idsToAckSet.contains(bindingID)) {       // O(1) → O(R+A) total
+                binding.routeWithAck(message, context);
+            } else {
+                binding.route(message, context);
+            }
+        }
+    }
+}
+```
+
+## Complexity
+| | Before | After |
+|---|---|---|
+| routeFromCluster | O(R×A) | O(R+A) |
+
+Where R=routing targets, A=ACK targets.
+At R=100 bindings, A=50 ACKs: 5,000 ops per message → 150 ops (33x improvement).
+Under sustained load of 10,000 msg/s this saves ~49.8M operations/second.
+
+## Context
+This is in the clustered message routing path — called when the broker receives a
+message with `HDR_ROUTE_TO_ACK_IDS` set, routing to multiple queues where some
+require acknowledgment. In large clustered deployments with many queues per address,
+R and A can be large.